Operate customer-facing Kubernetes clusters on your own GPU infrastructure
This tool is for platform teams that want to provide Kubernetes services to external customers using their own GPU servers. NVKE installs a management plane to create multiple independent workload clusters and manages the lifecycle of each cluster. Customers use their clusters with kubectl, while operators control the infrastructure through NVKE.
Preview stage considerations
NVKE is currently provided in Preview status and is not recommended for production workloads. Because the API, configuration, and behavior may change with each release, you must check the latest documentation. The management cluster itself does not support upgrade, scaling, or replacement features, and there are no backup and recovery procedures. API authentication is disabled by default and communication occurs over plain HTTP, so you must configure authentication and security settings directly before allowing external callers.
Cluster creation and lifecycle management
NVKE provisions machines through NVIDIA Infra Controller (NICo) and configures an independent control plane, etcd, and node pool for each workload cluster. When you request a cluster via nvkectl or the NVKE API, the infrastructure is prepared automatically. You can manage node pool creation, scaling, deletion, Kubernetes version upgrades, and node replacement. Each node joins with credentials after SPIFFE/SPIRE authentication using TPM, and etcd data encryption can be configured per cluster.
Local evaluation and actual installation path
To evaluate without GPU hardware, you can use the local quickstart. This requires a Linux host or VM with amd64 or arm64 architecture, at least 6 CPUs, 12 GiB of memory, and 100 GiB of disk, along with Docker and virtual TPM support. On macOS, you evaluate using a Colima virtual machine. To install on actual infrastructure, you must prepare machines, networks, node images, and registries directly, and set up access to NICo. In the current preview stage, NICo is the only compatible infrastructure provider.
Source build and development environment
To build from source, you must use Go 1.27.3 and GNU Make 4 or higher. Running the make build command generates 9 binaries in the bin/ directory. The build process does not start the management plane or create clusters, and the node agent runs only on Linux. Before submitting changes, you must run make preflight to perform checks on the modified modules.
Pre-adoption checks
NVKE is a self-hosted Kubernetes cluster factory, not a simple Kubernetes distribution or a Cluster API flavor. The management cluster and workload clusters are clearly separated, allowing the management plane to operate multiple customer clusters independently. FDE node images are provided for evaluation only and do not support data encryption (data-at-rest). Before adoption, you must verify the known limitations and the validated combinations and pinned dependencies through the platform and version reference documentation.
